foxclaw 0.1.0 → 0.1.2

package/python/vendor/hmr/engines/lifecycle.py

@@ -0,0 +1,264 @@
+"""
+Memory Lifecycle Engine - 记忆生命周期引擎
+HMR v1.5 新增
+
+解决问题：
+- compress_memories 需要手动调用 → 自动触发
+- 低价值记忆持续占用空间 → 自动 prune
+- 同类记忆堆积 → 自动 consolidate
+
+生命周期状态：
+  新生(fresh) → 活跃(active) → 衰退(fading) → 休眠(dormant) → 删除(pruned)
+                                                        ↘ 合并(consolidated)
+"""
+
+import threading
+from typing import List, Dict, Optional, Callable, Any
+from dataclasses import dataclass, field
+from datetime import datetime
+from ..core.models import MemoryObject
+
+
+@dataclass
+class LifecycleConfig:
+    """生命周期配置"""
+    # 触发条件
+    max_memories_per_type: int = 80       # 同类型超过此数触发压缩
+    prune_retrievability: float = 0.05    # SM-2 可提取性低于此值考虑删除
+    prune_min_age_days: int = 7           # 至少存在 7 天才允许 prune
+    prune_require_zero_access: bool = True  # 必须从未被访问才 prune
+
+    # 压缩参数
+    consolidation_batch: int = 20         # 每次压缩多少条
+    consolidation_keep_ratio: float = 0.3 # 压缩后保留原记忆的 30%（降低权重而非删除）
+
+    # 自动运行
+    auto_enabled: bool = True
+    check_interval_ingests: int = 10      # 每 10 次 ingest 检查一次（避免每次都跑）
+
+
+@dataclass
+class LifecycleReport:
+    """生命周期执行报告"""
+    timestamp: datetime = field(default_factory=datetime.utcnow)
+    pruned_count: int = 0
+    consolidated_count: int = 0
+    compressed_to: int = 0                # 压缩后生成了几条新记忆
+    checked_count: int = 0
+    reasons: List[str] = field(default_factory=list)
+
+    def summary(self) -> str:
+        parts = [f"检查 {self.checked_count} 条"]
+        if self.pruned_count:
+            parts.append(f"删除 {self.pruned_count} 条")
+        if self.consolidated_count:
+            parts.append(f"压缩 {self.consolidated_count} → {self.compressed_to} 条")
+        return "；".join(parts) if parts else "无操作"
+
+
+class MemoryLifecycleEngine:
+    """
+    记忆生命周期引擎
+
+    在 HMR 主引擎里，每 N 次 ingest 后自动调用：
+        lifecycle.check(triggered_by=new_memory)
+
+    核心决策流：
+        ┌──────────────┐
+        │  所有记忆     │
+        └──────┬───────┘
+               ↓
+        ┌──────────────┐   可提取性 < 0.05
+        │  衰退检测    │ ──────────────────→  pruned（删除）
+        └──────┬───────┘
+               │ 未到删除线
+               ↓
+        ┌──────────────┐   同类型 > max
+        │  数量检测    │ ──────────────────→  consolidate（压缩）
+        └──────┬───────┘
+               │ 数量正常
+               ↓
+              保留
+    """
+
+    def __init__(self, memory_fs, temporal_engine, config: LifecycleConfig = None):
+        self.memory_fs = memory_fs
+        self.temporal_engine = temporal_engine
+        self.config = config or LifecycleConfig()
+        self._lock = threading.Lock()
+        self._ingest_counter = 0
+        self._compress_fn: Optional[Callable] = None  # 注入 hmr.compress_memories
+
+    def register_compress_fn(self, fn: Callable):
+        """注入压缩函数（避免循环引用）"""
+        self._compress_fn = fn
+
+    def on_ingest(self, new_memory: MemoryObject) -> Optional[LifecycleReport]:
+        """
+        每次 ingest 后调用。
+        根据 check_interval_ingests 决定是否真正执行检查。
+        """
+        if not self.config.auto_enabled:
+            return None
+
+        with self._lock:
+            self._ingest_counter += 1
+            should_check = (self._ingest_counter % self.config.check_interval_ingests == 0)
+
+        if not should_check:
+            return None
+
+        # 异步执行，不阻塞 ingest
+        report = LifecycleReport()
+        t = threading.Thread(
+            target=self._run_check,
+            args=(new_memory.type, report),
+            daemon=True
+        )
+        t.start()
+        return report   # report 会在后台填充
+
+    def check_now(self, memory_type: Optional[str] = None) -> LifecycleReport:
+        """手动触发完整检查（同步）"""
+        report = LifecycleReport()
+        self._run_check(memory_type, report)
+        return report
+
+    def _run_check(self, memory_type: Optional[str], report: LifecycleReport):
+        """执行生命周期检查（在后台线程中运行）"""
+        try:
+            # 1. Prune：删除低价值记忆
+            self._prune(memory_type, report)
+
+            # 2. Consolidate：压缩数量过多的记忆
+            self._consolidate(memory_type, report)
+
+        except Exception as e:
+            report.reasons.append(f"生命周期检查失败: {e}")
+            print(f"[HMR Lifecycle] 错误: {e}")
+
+    def _prune(self, memory_type: Optional[str], report: LifecycleReport) -> List[str]:
+        """删除满足条件的低价值记忆"""
+        memories = self.memory_fs.list_memories(memory_type)
+        report.checked_count = len(memories)
+        pruned_ids = []
+        now = datetime.utcnow()
+
+        for mem in memories:
+            # 条件 1：存在足够长时间
+            age = (now - mem.created_at).days
+            if age < self.config.prune_min_age_days:
+                continue
+
+            # 条件 2：从未被访问（如果配置要求）
+            if self.config.prune_require_zero_access and mem.access_count > 0:
+                continue
+
+            # 条件 3：SM-2 可提取性极低
+            if mem.id in self.temporal_engine.sm2_states:
+                r = self.temporal_engine.sm2_states[mem.id].retrievability()
+            else:
+                r = mem.temporal_weight
+
+            if r < self.config.prune_retrievability:
+                # 执行删除
+                deleted = self.memory_fs.delete_memory(mem.id)
+                if deleted:
+                    pruned_ids.append(mem.id)
+                    report.pruned_count += 1
+                    report.reasons.append(
+                        f"删除 [{mem.type}]《{mem.title}》"
+                        f"（可提取性={r:.3f}，存在{age}天，访问0次）"
+                    )
+
+        if pruned_ids:
+            print(f"[HMR Lifecycle] 自动删除 {len(pruned_ids)} 条低价值记忆")
+
+        return pruned_ids
+
+    def _consolidate(self, memory_type: Optional[str], report: LifecycleReport):
+        """当同类型记忆超过阈值时自动压缩"""
+        if not self._compress_fn:
+            return
+
+        types_to_check = [memory_type] if memory_type else [
+            "execution", "reflection", "agent_memory"
+        ]
+
+        for mtype in types_to_check:
+            if not mtype:
+                continue
+            memories = self.memory_fs.list_memories(mtype)
+
+            if len(memories) <= self.config.max_memories_per_type:
+                continue
+
+            # 找出最老/最不活跃的一批来压缩
+            candidates = sorted(
+                memories,
+                key=lambda m: (m.temporal_weight, m.access_count)
+            )[:self.config.consolidation_batch]
+
+            if len(candidates) < 3:
+                continue
+
+            print(f"[HMR Lifecycle] [{mtype}] 记忆数={len(memories)}，自动压缩 {len(candidates)} 条...")
+
+            try:
+                compressed = self._compress_fn(
+                    memory_ids=[m.id for m in candidates]
+                )
+                if compressed:
+                    report.consolidated_count += len(candidates)
+                    report.compressed_to += 1
+                    report.reasons.append(
+                        f"[{mtype}] {len(candidates)} 条 → 1 条《{compressed.title}》"
+                    )
+            except Exception as e:
+                print(f"[HMR Lifecycle] 压缩失败: {e}")
+
+    def get_lifecycle_stats(self) -> Dict[str, Any]:
+        """获取记忆生命周期统计"""
+        all_memories = self.memory_fs.list_memories()
+        now = datetime.utcnow()
+
+        stats = {
+            "total": len(all_memories),
+            "by_state": {"fresh": 0, "active": 0, "fading": 0, "dormant": 0},
+            "by_type": {},
+            "at_risk": []  # 即将被 prune 的记忆
+        }
+
+        for mem in all_memories:
+            # 获取可提取性
+            if mem.id in self.temporal_engine.sm2_states:
+                r = self.temporal_engine.sm2_states[mem.id].retrievability()
+            else:
+                r = mem.temporal_weight
+
+            # 分类状态
+            age = (now - mem.created_at).days
+            if age < 1:
+                state = "fresh"
+            elif r > 0.7:
+                state = "active"
+            elif r > 0.3:
+                state = "fading"
+            else:
+                state = "dormant"
+
+            stats["by_state"][state] += 1
+            stats["by_type"][mem.type] = stats["by_type"].get(mem.type, 0) + 1
+
+            # 标记风险记忆
+            if r < self.config.prune_retrievability * 2 and mem.access_count == 0:
+                stats["at_risk"].append({
+                    "id": mem.id,
+                    "title": mem.title,
+                    "type": mem.type,
+                    "retrievability": round(r, 3),
+                    "age_days": age
+                })
+
+        stats["at_risk"] = stats["at_risk"][:10]  # 最多显示 10 条
+        return stats